For the technical evaluator
Where the second goes.
A voice agent is judged on one number: the gap between a caller finishing a sentence and hearing the first word back. Everything below exists to keep that gap under a second and a half, and to keep it there in minute twelve.
- 01
Endpointing
Caller stops speaking → we know they stopped
semantic, not a fixed silence timer
- 02
Transcription
Speech → text
streaming, partials consumed as they arrive
- 03
Reasoning
Text → the next thing to say
first token streamed, not awaited whole
- 04
Synthesis
Text → audio
first chunk plays before the sentence is finished
- 05
Carriage
Audio → the caller's ear
warm connections, pooled per region
Building an agent
Write it in English, or draw the call.
Both produce the same agent. A prompt agent is one brief and a set of tools; the workflow canvas is the same call laid out step by step, for when the path has to be guaranteed rather than reasoned about.
The pieces a call is built from
- greeting
- conversation
- collect details
- look up a document
- call an API
- set a variable
- condition
- send an SMS
- send an email
- send a WhatsApp
- transfer to a number
- transfer to another agent
- goodbye
What keeps it honest
- Knowledge retrieval
- Answers are retrieved from your uploaded documents at call time, so editing the file changes the answer without republishing the agent.
- Webhook tools
- The agent calls your endpoints mid-conversation to check availability, look up a record, or write a booking — and waits for the real answer.
- The caller-number rule
- The caller’s number is supplied by the carrier, never by the caller. An agent cannot be talked into acting on a number the caller reads out.
- Structured outputs
- Every call ends with the fields you asked for — intent, outcome, entities — extracted from the transcript rather than typed by anyone.
After the call
Recorded in stereo, so you can hear who talked over whom.
Two channels, one timeline
Caller on one channel, agent on the other. Interruptions and overlaps are audible rather than inferred, which is the only way to debug a turn that went wrong.
Scenario testing
Run a written scenario against a draft agent and diff the outcome before it ever touches a real line. Nothing publishes until it passes.
Versioning
Publishing snapshots the whole agent — prompt, tools, knowledge, voice. Any version restores.
Bring your own, at every layer
- Carriers
- Twilio · Plivo · SIP trunk
- Speech to text
- Deepgram · Soniox · Sarvam
- Reasoning
- OpenAI · Anthropic · Google · Groq
- Speech
- ElevenLabs · Cartesia